More screenshots
Works with the AI Agents you already use
Mnemosyne Pro — Production Agent Memory
The tested production setup behind 409/500 on LongMemEval-S: prove reranked recall is live, wire Claude Code and Codex, migrate without loss.
$19
Mnemosyne Pro — Production Agent Memory
Example session with this skill installed
Check my Mnemosyne setup in this project and tell me whether the reranked retrieval path is really active.
- Read your context and instructions
- Compiled the mnemosyne pro
Setup check: PASS, reranked recall is live
Real output of the bundled check on a project pinned to mnemosy-ai 2.0.0 with the provisioned model cache:
OK Node 26.12.1
OK package.json pins mnemosy-ai to 2.0.0
OK installed mnemosy-ai 2.0.0
OK @huggingface/transformers pinned to 3.8.1
OK overrides["@huggingface/transformers@3.8.1"].sharp = 0.35.4
INFO sqlite-vec loads; indexMode 'auto' uses it (still an exact search)
INFO hnswlib-node does not load (absent, or its native binding is not built); do not set indexMode 'hnsw'
OK embedder weights (34014426 bytes)
OK embedder vocabulary (231508 bytes)
OK reranker weights (23143499 bytes)
OK reranker vocabulary (231508 bytes)
all required checks passed
Connects securely to your tools. The creator never sees your data.
About this skill
Production memory, proven on. Mnemosyne Pro is the tested setup we run around the free, MIT-licensed Mnemosyne engine (mnemosy-ai 2.0.0): the exact configuration behind 409 of 500 on LongMemEval-S, and the checks that prove your agent is really on that path.
Why it matters
A plain install has no models. Without them, hybrid recall silently falls back to keyword search: no error, nothing in the logs. This skill provisions the two pinned local models (after asking you), verifies them by SHA-256 offline, and proves which retrieval path is live.
What's inside
- The 409/500 configuration. Every retrieval and reader setting, how each was chosen on a dev split, the untouched-holdout result (82.25% on 400 questions), eight honest caveats, and a type-checked TypeScript module that reproduces it in your app.
- Provision & prove. Pinned embedder and reranker, offline SHA-256 verification, and a one-line retrieval-path check.
- Claude Code & Codex recipes. Server entries generated by their own CLIs, one database per project, pinned versions, an explicit model cache, and the right tool set per agent role.
- Upgrade without loss. A staged rollout from 1.x with instant rollback, plus tested preview, apply and undo imports from other memory-store exports.
- Measure your own data. Reproduce the published numbers, run the accuracy driver for $0 before paying for a real run, and score your own corpus.
- 21 tested fixes. Symptom, cause and fix with the exact error text.
- Two offline scripts. A model verifier and a read-only project check.
How it's built
Every command was run against the released mnemosy-ai@2.0.0, and the outputs shown are what it printed. Anything we could not run, such as a live Cursor session, is marked where it appears. The skill asks before installing packages, downloading models, editing host configuration, applying a migration or calling a paid API.
Who it's for
Teams putting agent memory into production who need recall quality they can prove, not assume.
New to Mnemosyne? Start with the free skill, Mnemosyne Memory (official).
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.
- Not a hosted service: the engine, database and models stay on your machine.
- Not a licence or activation step: every feature it uses is in the free MIT package.
- Does not change two disclosed engine targets: recall p95 at 1M records is 139 to 223 ms against a 50 ms goal, and contextual reranking shipped 0.82 points under its gate.
- The 409/500 is one selected configuration on a public benchmark, judged by an LLM, with a reader model you pay for separately (full run cost $9.11 at list prices). Measure on your own data first.
- Imports read export files; they are not a live sync.
- The Cursor entry follows Cursor's documentation and was not run in Cursor.
- Email support only, no response-time commitment.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 20 days ago
- Accounts
Hugging Face
GitHub
Openrouter
- Software
Node.js 22.16+npmnpxmnemosy-ai 2.0.058 MB free disk spaceBash
Python
Gittscmnemosy
The disk space holds the pinned models. Your agent must be an MCP host; Claude Code and Codex are tested.
Permissions
Allowed Hosts
File Scopes
Every install, model download, config write, migration apply and paid call happens only after the user agrees. npm and Hugging Face are contacted once for the package and the pinned models. OpenRouter is used only if you opt in to paid reader/judge runs with your own key. GitHub and dataset hosts only for the optional benchmark steps. Both bundled scripts run offline and write nothing. No telemetry.